WHERE AGENTIC AI BREAKS HERE
Disparate-impact bias at population level
Each candidate's score reads defensible. The pattern across protected classes, taken over a sufficient sample, fails an adverse-impact analysis.
Prompt injection through candidate inputs
CVs, free-text answers, and recorded responses are injection surfaces. Nudged agentic AI applies inconsistent weighting it cannot defend.
Hallucinated competency signals
The agent reports a high-confidence competency signal with no evidentiary trail behind it. The reviewer treats it as ground truth.
Assessment agent scores the candidate and produces competency signals
CVs, recorded responses, structured exercises, and free-text inputs all read inside the autonomous loop.
Disseqt tests against 84+
jailbreak techniques and
bias-probe library
Demographic-perturbation tests, adversarial candidate profiles, and the 67+ validator suite all run as a standard pre-release gate.
HR reviewer sees confidence score and root-cause analysis on flagged outputs
The reviewer reads what triggered the flag, the population context, and the specific score that needs human override.
EEOC, EU AI Act, and Local Law 144 disclosure pack per workflow run
Bias-audit and disclosure-ready artefacts assembled from live assessments, ready for HR and compliance review on demand
Per-candidate bias and consistency scoring
Every assessment output evaluated against bias signals and population context before it reaches HR.
Measurable bias posture release-over-release
Demographic-perturbation tests run on every model update, with the testing record to defend it.
Disclosure-ready evidence pack
EEOC, EU AI Act, and NYC Local Law 144 artefacts generated from live assessments, ready for legal or regulator review.
One pattern, adjacent workflows
The same assurance shape reused across hiring screen, promotion review, and adjacent talent workflows.
